Multi-Agent Pathfinding with Hierarchical Evolutionary Hueristic A*

Abstract

Multiagent pathfinding (MAPF) problem is an important topic to various domains including video games, robotics, logistics, and crowd simulation. The goal of a pathfinding algorithm for MAPF is to find the shortest possible path for each agent without collisions with other agents. Search is among the most fundamental techniques for problem solving, and A* is the best known heuristic search algorithm. While A* guarantees to find the shortest path using a heuristic function, it cannot handle the large scale and many uncertainties in MAPF. The main challenge of MAPF is the scalability. The problem complexity grows exponentially as both the size of environments and the number of autonomous agents increase, which becomes more difficult for A* to compute results in real time under the constraints of memory and computing resources. To overcome the challenges in MAPF, distributed approaches are introduced to reduce the computational time and complexity. Contrast to centralized approaches, which use a single controller to determine every move of all agents, distributed approaches allow each agent to search for its own solution. Distributed MAPF algorithms need to refine solutions for all agents that are collision-free. The algorithm should lead agents to take another path, or standby on the same node at the moment, to avoid conflicts between any two paths. Under the circumstances, an optimal solution is no longer simply finding the shortest path for each agent. Instead, it should contain a collision-free path for every agent, with the lowest makespan, which is the number of time steps required for all agents to reach their target. However, minimizing the makespan and the sum of cost for all agents is a NP-hard problem. Given MAPF problems often require to be solved in real time with limited resources, minimizing only the makespan is a more practical approach.To achieve accurate search and high scalability, a MAPF algorithm must fulfill the following requirements: 1) it is capable to compute collision-free paths for all agents; 2) it can provides an accurate priority decision mechanism to ensure solution optimality; and 3) it should maintain the successful rate to obtain a solution as the number of agents increases. In this paper, we proposed a novel hierarchical pathfinding technique named Multi-Agent Hierarchical Evolutionary Heuristics A* (MA-HEHA*). Our contributions in this paper are: 1) we propose MA-HEHA* that can identify bottleneck areas to reduce collisions in abstract search; 2) our algorithm evolves heuristic functions by itself to avoid potential conflicts during local search; 3) we prove that MA-HEHA* maintain high successful rate when the scalability is high; 4) we evaluate MA-HEHA* on different types of MAPF problems to show its effectiveness. Our experiment results show that ourMA-HEHA* can efficiently solve large scale MAPF problems compared to traditional MAPF approaches.